Accessibility setting

Section: New Results

Dimension free principal component analysis

Participants : Olivier Catoni, Ilaria Giulini.

In a work in progress, Ilaria Giulini, as part of her PhD studies, proved the following dimension free inequality, related to Principal Component Analysis in high dimension. Given an i.i.d. sample $X_i$, $1\le i\le n$ of vector valued random variables $X_i\in {𝐑}^d$, there exists an estimator $\widehat{N}$ of the quadratic form $N\left(\theta \right)=𝐄\left({\langle \theta ,X\rangle }^2\right)$ such that for any $n\le {10}^{20}$, with probability at least $1-2ϵ$, for any $\theta \in {𝐑}^d$,

where

where $G=𝐄\left(X{X}^{\top }\right)$ is the Gram matrix and where ${\displaystyle \kappa =sup\left\{\frac{𝐄\left({\langle \theta ,X\rangle }^4\right)}{𝐄{\left({\langle \theta ,X\rangle }^2\right)}^2},\theta \in {𝐑}^d\setminus \mathrm{𝐊𝐞𝐫}\left(G\right)\right\}}$ is some kurtosis coefficient. This result proves that the expected energy in direction $\theta$ can be estimated at a rate that is independent of the dimension of the ambient space ${𝐑}^d$. It is obtained using PAC-Bayes inequalities with Gaussian parameter perturbations. The same bound holds in a Hilbert space of infinite dimension, opening the possibility of a rigorous mathematical study of kernel principal component analysis of random data, where the data are represented in a possibly infinite dimensional reproducing kernel Hilbert space.